Research And Design Of Semantic SLAM With Dynamic Environment

The posture changes of autonomous mobile robots are estimated by the technology of Simultaneous Localization and Mapping,and it also used to construct the 3D environment map which contains spatial geometry.In recent years,typical SLAM solutions have achieved real-time location tracking and sparse map construction in a large range of static scenarios.However,the traditional SLAM scheme still has the following problems: First,in the dynamic environment,it is impossible to avoid the drift error caused by moving objects such as pedestrians,so its application in reality is subject to certain limitations.Second,robots have poor performance of understand surroundings if just only to generate sparse spatial maps containing geometric information.This study focusses on the positioning of indoor robots and the reconstruction of environment under dynamic scenes.The estimation error caused by moving objects is solved that the semantic information of image is extracted by deep learning network in the process of position estimation.And it can generate static 3D environment maps.The details of the study are as follows:1.A method of image motion area detection based on semantic segmentation network is studied in this research to divide the potential motion objects in indoor environment.The performance of semantic segmentation network is verified by the experiment designed in this study to have a theoretical foundation that can design the feature point filtering algorithm in visual odometer.2.A feature point selection algorithm that combines semantic information and dynamic target detection is proposed in this study to solve the influence of dynamic objects on SLAM positioning.Based on public datasets TUM and measured scene test datasets for locating and tracking,which calculate the relative position error and absolute trajectory error to evaluate the performance between the improved SLAM algorithm in this study and traditional ORB-SLAM2 algorithm.These have similar performance of positioning accuracy under static environment through the analysis of experimental results,but the average square root errors of the global estimated trajectory and relative position error of the improved SLAM is reduced by 96.51% and60.10% under dynamic scene.3.A map generation algorithm of point cloud dense in dynamic environment is proposed in this study.The problem which lacks map information can be solved as a map generated by SLAM.The results based on the RGB-D test data in different indoor scenarios show that the algorithm has a stable and effective performance under complex environment.Finally,an octomap is generated by the point cloud map to improve the robustness and efficiency of map construction including SLAM's 3D environment and dynamic environment.The system of Ubuntu 16.04 is used to build the experimental software environment which used to verify the effectiveness of these algorithms proposed in this study,and the experiment has completed based on public dataset of TUM and measured data collected under the scene established by laboratory.The EVO tools use to analysis the experimental result,and the algorithm of improved SLAM in this research has a perfect performance of robustness and effectiveness under dynamic environment.